320 likes | 494 Views
bipolate : A Stata command for bivariate interpolation with particular application to 3D graphing. Joseph Canner, MHS Xuan Hui, MD ScM Eric Schneider, PhD Johns Hopkins University Stata Conference Boston, MA August 1, 2014. Background. Educational outcome study Continuous outcome
E N D
bipolate: A Stata command for bivariate interpolation with particular application to 3D graphing Joseph Canner, MHS Xuan Hui, MD ScM Eric Schneider, PhD Johns Hopkins University Stata Conference Boston, MA August 1, 2014
Background • Educational outcome study • Continuous outcome • Two categorical (1-5) predictors • Panel data • 8 time periods • 285 observations per time period • Researcher desired 3D plot of outcome versus two predictors
Solution #2: surface* . surface CommunicationNEWEduNEW COMPSS, xlabel(1/5) ylabel(1/5) * by Adrian Mander, available from SSC
collapse first . collapse (mean) mCOMPSS=COMPSS, by(EduNEWCommunicationNEW) . surface CommunicationNEWEduNEWmCOMPSS, xlabel(1/5) ylabel(1/5)
SAS Solution proc g3grid data=a out=b ; grid EduNEW*CommNEW=mCOMPSS/ axis1=1 to 5 by 0.1 axis2=1 to 5 by 0.1; run;
SAS Solution (cont’d) proc g3d data=b; plot EduNEW*CommNEW=mCOMPSS/ xticknum=5 yticknum=5 grid; run;
Stata Conference 2013 Wishes and grumbles session: no plans to implement 3D graphing
SAS PROC G3GRID • Interpolation options: • <default>: biquintic polynomial • PARTIAL: use splines for derivatives • NEAR=n: number of nearest neighbors (default=3) • SPLINE: bivariate spline • SMOOTH=numlist: smoothed spline • JOIN: linear interpolation
Bivariate interpolation • SAS G3GRID default • Akima, Hiroshi (1978), "A Method of Bivariate Interpolation and Smooth Surface Fitting for Irregularly Distributed Data Points," ACM Transaction on Mathematical Software, 4, 148-159. • Fortran 77 originally in the NCAR library; Fortran 77 and Fortran 90 versions freely available on the web
History “The original version of BIVAR was written by Hiroshi Akimain August 1975 and rewritten by him in late 1976. It was incorporated into NCAR's public software libraries in January 1977. In August 1984 a new version of BIVAR, incorporating changes described in the Rocky Mountain Journal of Mathematics article cited below, was obtained from Dr. Akima by Michael Pernice of NCAR's Scientific Computing Division, who evaluated it and made it available in February, 1985.” Ref: Hiroshi Akima, On Estimating Partial Derivatives for Bivariate Interpolation of Scattered Data, Rocky Mountain Journal of Mathematics, Volume 14, Number 1, Winter 1984.
Algorithm summary • XY plane divided into triangular cells • Bivariate quintic polynomial in X and Y fitted to each triangular cell • Coefficients are determined by continuity requirements and by estimates of partial derivatives at the vertices and along triangle edges
Algorithm features • Invariant to certain transformations: • Rotation of XY coordinate system • Linear scale transformation of the Z axis • Tilting of the XY plane
Algorithm features (cont’d) • Interpolating function and first-order partial derivatives are continuous • Local method: change in data in one area does not effect the interpolating function in another area • Gives exact results when all points lie in a plane
bipolate command Syntax: bipolatexvaryvarzvar [if] [in] [using] [, options]
bipolate options • method: interpolation or filling • xgrid, ygrid: specify x-axis and y-axis values to use for interpolation • fillusing: specify data set to use for filling • collapse: how to handle multiple values of z at a given x and y • saving: save the resulting data to set to disk
Use of bipolate . bipolateCommunicationNEWEduNEWCOMPSS, xgrid(1(0.1)5) ygrid(1(0.1)5) method(interp) saving(test_bip) . use test_bip, clear . surface EduNEWCommunicationNEWCOMPSS_mod
Remaining puzzles • Why are there small differences between interpolated values? • SAS: “This default method is a modification of that described by Akima (1978)” • Re-orienting axis
Future plans • Make available on SSC within a few weeks • Test other data sets • Testing and debugging by Stata community
Possible Future Enhancements • Implement partial and near options • Implement scale/noscale option • Implement spline and smooth spline interpolation • See if Mata has functions that can reproduce the algorithm more compactly